A monorepo showcasing modern cloud-native and AI-powered workflows. Built on Crossplane v2 for platform API abstractions and FluxCD for GitOps automation. Serves as a playground for exploring the intersection of infrastructure-as-code, AI agents, and Kubernetes-native tooling.

Write-ups and demos: Repo's Wiki · demo/

Logo	Name	Description	Project Version
	GKE	Google Kubernetes Engine is Google Cloud's managed Kubernetes service	1.34.1
	Gateway API	Kubernetes Gateway API is a collection of resources that model service networking in Kubernetes, providing expressive, extensible, and role-oriented interfaces.	v1.5.1
	kgateway	Kubernetes gateway for AI services, providing a standardized way to connect applications with AI capabilities within the cluster.	v2.2.2
	Agent Gateway	Gateway Dataplane for AI workloads (MCP, A2A)	v1.0.1
	Crossplane	Open-source Kubernetes add-on that lets you provision and manage cloud infrastructure and external services using Kubernetes-style APIs and declarative configuration.	v2.2.0
	kagent	Kubernetes-native AI agent framework that enables the deployment and management of AI agents within Kubernetes clusters.	v0.7.21
	FluxCD	GitOps toolkit for Kubernetes that keeps clusters in sync with configuration sources and automates deployments.	v2.8.3
	LitmusChaos	Cloud-native chaos engineering framework for Kubernetes that helps teams find weaknesses in their deployments through controlled chaos experiments.	v3.27.0

Three-cluster hub-and-spoke fleet, provisioned in order:

kind (local bootstrap)
  └─ provisions → GKE control-plane   (Crossplane + Flux + platform services)
                    └─ provisions → GKE apps-dev   (tenant workloads)

• Crossplane v2 handles GKE cluster provisioning (no claims — direct namespace-scoped XRs)
• Flux GitOps manages everything on each cluster once bootstrapped
• GitHub Actions bootstraps Flux on new GKE clusters, triggered by Flux notifications

The provisioning pipeline is driven by a Claude-powered agentic loop that monitors, diagnoses, and fixes the cluster fleet without human intervention:

task agentic:deploy
        │
        ├─ bootstrap-control-plane-cluster.sh  (background)
        │
        └─ phase loop  [bootstrap → control-plane → apps-dev]
             ├─ collect kubectl state
             ├─ phase-checker agent  →  healthy | wait | diagnose | teardown
             └─ diagnostics agent   →  fix_forward | teardown | escalate
                  fix_forward: edits manifests, commits to develop, waits for Flux to reconcile

Two Claude sub-agents drive the loop:

• phase-checker — evaluates cluster state against healthy criteria for the current phase; returns a structured JSON verdict
• diagnostics — investigates failures, reads the mission context, makes targeted manifest edits, and returns a commit-ready fix

All cluster changes go through git → Flux. Direct kubectl writes are blocked by a PreToolUse hook. The orchestrator tracks error signatures and escalates after 3 identical failures or 2 full teardown cycles.

→ Full reference: docs/orchestrator.md

• Access to a GCP account with sufficient permissions
• Tools: gcloud, flux, kubectl, task, uv, Claude Code (claude CLI)
• GitHub organisation or personal account

See docs/github-integration.md for GCP OIDC / GitHub Actions setup.

# Scripted deploy (no AI)
task setup:deploy

# Agentic deploy — same goal, AI-monitored and self-fixing
task agentic:deploy

# Resume agentic deploy from a specific phase (cluster already exists)
task agentic:resume PHASE=control

# Tear everything down
task setup:cleanup

→ Diagnostics, load tests, and performance experiments: docs/operations.md